Simulation of site‐scale water fluxes in desert and natural oasis ecosystems of the arid region in Northwest China

Abstract

The study of water fluxes is important to better understand hydrological cycles in arid regions. Data‐driven machine learning models have been recently applied to water flux simulation. Previous studies have built site‐scale simulation models of water fluxes for individual sites separately, requiring a large amount of data from each site and significant computation time. For arid areas, there is no consensus as to the optimal model and variable selection method to simulate water fluxes. Using data from seven flux observation sites in the arid region of Northwest China, this study compared the performance of random forest (RF), support vector machine (SVM), back propagation neural network (BPNN), and multiple linear regression (MLR) models in simulating water fluxes. Additionally, the study investigated inter‐annual and seasonal variation in water fluxes and the dominant drivers of this variation at different sites. A universal simulation model for water flux was built using the RF approach and key variables as determined by MLR, incorporating data from all sites. Model performance of the SVM algorithm (R2 = 0.25–0.90) was slightly worse than that of the RF algorithm (R2 = 0.41–0.91); the BPNN algorithm performed poorly in most cases (R2 = 0.15–0.88). Similarly, the MLR results were limited and unreliable (R2 = 0.00–0.66). Using the universal RF model, annual water fluxes were found to be much higher than the precipitation received at each site, and natural oases showed higher fluxes than desert ecosystems. Water fluxes were highest during the growing season (May–September) and lowest during the non‐growing season (October–April). Furthermore, the dominant drivers of water flux variation were various among different sites, but the normalized difference vegetation index (NDVI), soil moisture and soil temperature were important at most sites. This study provides useful insights for simulating water fluxes in desert and oasis ecosystems, understanding patterns of variation and the underlying mechanisms. Besides, these results can make a contribution as the decision‐making basis to the water management in desert and oasis ecosystems.

Evaluation of water flux predictive models developed using eddy-covariance observations and machine learning: a meta-analysis

“…6) appear to be insignificant. This seems to be related to the fact that in large-scale water flux simulations, the sites of similar PFTs are selected such as for modeling multiple forest sites across Europe (Van Wijk and Bouten, 1999) which focus on "forest" and multiple grassland sites across arid northern China (Xie et al, 2021;Zhang et al, 2021) which focus on "grassland", rather than mixing different PFT types to train models as is done in machine learning modeling of carbon fluxes (Zeng et al, 2020). In terms of the timescales of the models, the 4 d, 8 d, and monthly scales appear to correspond to higher accuracy compared to the halfhourly and daily scales.…”

“…Currently, there are three main approaches for simulation and spatial and temporal prediction of ET: (i) physical models based on remote sensing, such as surface energy balance models (Minacapilli et al, 2009;Wagle et al, 2017), the Penman-Monteith equation (Mu et al, 2011;Zhang et al, 2010), and the Priestley-Taylor equation (Miralles et al, 2011); (ii) process-based land surface models, biogeochemical models, and hydrological models (Barman et al, 2014;Pan et al, 2015;Sándor et al, 2016;Chen et al, 2019); and (iii) the observation-based machine learning modeling approach with in situ eddy-covariance (EC) observations of water flux (Jung et al, 2011;Li et al, 2018;Van Wijk and Bouten, 1999;Xie et al, 2021;Xu et al, 2018;Yang et al, 2006;Zhang et al, 2021). For remote-sensing-based physical models and process-based land surface models, some physical processes have not been well characterized due to the lack of understanding of the detailed mechanisms influencing ET under different environmental conditions.…”

Evaluation of water flux predictive models developed using eddy covariance observations and machine learning: a meta-analysis

“…6) appear to be insignificant. This seems to be related to the fact that in large-scale water flux simulations, the sites of similar PFTs are selected such as for modeling multiple forest sites across Europe (Van Wijk and Bouten, 1999) which focus on "forest" and multiple grassland sites across arid northern China (Xie et al, 2021;Zhang et al, 2021) which focus on "grassland", rather than mixing different PFT types to train models as is done in machine learning modeling of carbon fluxes (Zeng et al, 2020). In terms of the timescales of the models, the 4 d, 8 d, and monthly scales appear to correspond to higher accuracy compared to the halfhourly and daily scales.…”

“…Currently, there are three main approaches for simulation and spatial and temporal prediction of ET: (i) physical models based on remote sensing, such as surface energy balance models (Minacapilli et al, 2009;Wagle et al, 2017), the Penman-Monteith equation (Mu et al, 2011;Zhang et al, 2010), and the Priestley-Taylor equation (Miralles et al, 2011); (ii) process-based land surface models, biogeochemical models, and hydrological models (Barman et al, 2014;Pan et al, 2015;Sándor et al, 2016;Chen et al, 2019); and (iii) the observation-based machine learning modeling approach with in situ eddy-covariance (EC) observations of water flux (Jung et al, 2011;Li et al, 2018;Van Wijk and Bouten, 1999;Xie et al, 2021;Xu et al, 2018;Yang et al, 2006;Zhang et al, 2021). For remote-sensing-based physical models and process-based land surface models, some physical processes have not been well characterized due to the lack of understanding of the detailed mechanisms influencing ET under different environmental conditions.…”